My thoughts on Gas Town after 10,000 hours of Claude Code

15 points by deevus

spc476

It has probably hyperbole but there is no way the author has 10k hours of Claude Code---there are only 8,760 hours in a given year. Also, I saw it a while ago---the end game is the source code isn't important anymore.

deevus

I clearly haven't slept.
kevinc

Comment removed by author

doctor_eval

I just refuse to consider vibecoding as “vibe engineering”.

I consider engineering to the the application of scientific principles to the development of an artefact. This is no such thing. It’s not remotely scientific. It’s in the name!

This is not a value judgement. Writing a novel is not generally considered engineering, either. But Vibecoding by definition does not care how it’s done, and engineering, by most reasonable definitions, does.

Related,

I’ve never seen the code, and I never care to, which might give you pause. - Steve Yegge

I feel like there is going to be a reckoning in a year or two where all the shocking security shortcuts and lack of care that I’ve witnessed from Claude Code start to find their way into malign hands. If it hasn’t already.

deevus

Are you referring specifically to Gas Town, or agentic code generation in general?

The way I use Claude Code, I am using engineering principles, I’m just not writing 100% of the code. I guess what I’m saying is don’t throw out the baby with the bath water.
- doctor_eval
  
  Mostly Gas Town.
  
  I think Claude assisted coding is still engineering, as long as an engineer is still making decisions and all code is reviewed. I use Claude myself and it has made me more productive on some tasks, especially smaller changes, but I’ve learned to pick and choose what I get it to do.
  
  But that’s not vibecoding, which as I understand it is epitomised by Yegge’s comment that he’s “never seen the code”.
  
  Frankly, I think it’s absolutely irresponsible to publish a product with code you’ve never read. My guess is that it’s all going to come crashing down on folks. Especially people who are pushing out production code based on this stuff.
  
  LLMs are amazing technology, but the problem has never been writing the code, but understanding what needs to be done. If you spend just as much time writing English as you do writing code, I’m not sure what the benefit is.
rvt

I keep saying a lot of the agentic trends right now is humans figuring stuff out. None of it needs to be accepted as canon, and we can all take it as an interesting exploration of constraints (one of them being $$ to spin up lots of Claudes).

Opinion is good. The unfortunate thing is when BigCo et al. drop <Product Name / Protocol> and folks adopting it assuming it's a linear progression / improvement on past things. Instead we're very much in a breadth-first exploration state now IMO.
DanOpcode

Sounds like a fun program to try.

But I wonder how does the author use Claude Code for pair programming? Would be interesting with some more description of that workflow
- deevus
  I alluded to that in this post: https://simonhartcher.com/posts/2025-12-31-give-your-agentic-processes-a-name/
  
  Short version:
  
  Try using the brainstorming skill from https://github.com/obra/superpowers . This gives you way more insight into what is going to be delivered. The agent does all the coding while you watch and comment. With the executing-plans skill it will stop for feedback after each batch of tasks.
  
  Use the above with what I've called "test driven pair programming". Agent writes the tests, you write the implementation. This is very fun!
  - DanOpcode
    
    Thanks! I will check it out!